Cholesterol Levels and Risk of Hemorrhagic Stroke

Introduction

Hypercholesterolemia has been well documented as a modifiable risk factor for ischemic stroke.¹ Currently, lipid-lowering therapy with statins is widely used for patients with ischemic stroke.² However, concerns have been raised about the accompanied risk of hemorrhagic stroke, mainly including intracerebral hemorrhage (ICH) and subarachnoid hemorrhage (SAH), which may be attributed to decreasing serum cholesterol concentrations.^3,4

In the early Japanese studies, the inverse relationship between serum cholesterol level and increased risk of hemorrhagic stroke was first revealed.⁵ Later, in the Multiple Risk Factor Intervention Trial, the risk of fatal ICH was found to be 3× higher in those with total serum cholesterol (TC) <4.13 mmol/L than in those with values higher than that.⁶ In the collaborative analysis of 12 Asian cohorts, a 27% increase in risk of hemorrhagic stroke was shown for a 0.6 mmol/L decrease in cholesterol concentrations.⁷ However, conclusions were not consistent between studies. In the Korea Medical Insurance Corporation Study, low TC was not shown to be an independent risk factor for hemorrhagic stroke.⁸ In the post hoc analysis of Stroke Prevention by Aggressive Reduction in Cholesterol Levels (SPARCL) trial, although hemorrhagic stroke was more frequent in individuals treated with atorvastatin, its relations with concentrations of TC or low-density lipoprotein cholesterol (LDL-C) were not detected.³

Recent meta-analyses suggested no evidence that statins were associated with ICH.^9,10 However, stratified cholesterol levels were not explored. Hence, we undertook this meta-analysis, aiming to investigate the relationships between different categories of cholesterol and risk of hemorrhagic stroke.

Methods

Search Strategy

Two investigators (X.W. and Y.D.) independently searched PubMed and Embase for prospective studies examining the association between cholesterol and the risk of hemorrhagic stroke. Three main categories of cholesterol, including TC, high-density lipoprotein cholesterol (HDL-C), and LDL-C, were investigated, respectively. Hemorrhagic stroke mainly included ICH and SAH. The search was limited to studies published between 1980 and January 2013. The language was restricted to English. The detailed search strategy is given in Methods in the online-only Data Supplement.

Selection Criteria

To be included, studies had to have a prospective design (prospective cohort, or nested prospective case-control study), and investigate the association between cholesterol (TC, HDL-C, or LDL-C) and the risk of hemorrhagic stroke (ICH, SAH, or both). The authors should report effect estimates (risk ratio [RR], hazard ratio, or odds ratio) and 95% confidential intervals (CIs) for >3 categories of cholesterol concentrations, for the comparison between low and high concentrations, or for per 1 mmol/L increment. The detailed search strategy is provided in Methods in the online-only Data Supplement.

Data Extraction and Quality Assessment

Three authors (X.W.,Y.D., and X.Q.Q.) extracted the data in standardized data-collection forms, and 2 authors (X.W. and Y.D.) assessed the study quality. Three degrees for adjustment of confounders were defined. The Newcastle–Ottawa Scale was used to evaluate the methodological quality.¹¹ Details are described in Methods in the online-only Data Supplement.

Statistical Analysis

The RRs and associated 95% CIs were considered as the effect sizes. We first used the random-effects model to calculate summary RRs and 95% CIs for the high versus low levels of cholesterol. Then dose–response analyses were conducted to estimate the RRs and 95% CIs for per 1 mmol/L increment of cholesterol concentration. Heterogeneity was mainly assessed by the I² statistic.¹² We considered low, moderate, and high I² values to be 25%, 50%, and 75%, respectively.¹² Subgroup analyses and sensitivity analyses were also performed. A potential nonlinear relationship between cholesterol level and hemorrhagic risk was explored.¹³ Inter-rater reliabilities were calculated by Cohen κ statistics, with 5 levels of agreement, namely poor (κ=0.00–0.20), fair (κ=0.21–0.40), moderate (κ=0.41–0.60), good (κ=0.61–0.80), and very good (κ=0.81–1.00).¹⁴ The Egger test was used to assess publication bias.¹⁵ All statistical analyses were performed with the STATA software (version 12.0; Stata Corporation, College Station, TX). A threshold of P<0.1 was used to decide whether heterogeneity or publication bias was present.^15,16 In other ways, P values were 2 sided, with a significance level of 0.05. Detailed data are provided in Methods in the online-only Data Supplement.

Results

Literature Search

The results of study-selection process were shown in Figure 1. The initial search produced 453 studies from PubMed and 665 articles from Embase. After exclusion of duplicates and irrelevant studies, 97 potentially eligible studies were selected. After detailed evaluations, 23 studies were selected for final meta-analysis. A manual search of reference lists of these studies did not yield any new eligible study. Agreement on selection of studies between 2 assessors was very good (κ=0.87).

• Download figure
• Download PowerPoint

Study Characteristics

Twenty-three were included consisting of 19 prospective cohort studies,^8,17–34 and 4 nested case-control studies,^35–38 involving 1 430 141 participants with 7960 (5.6%) hemorrhagic stroke events (Table I in the online-only Data Supplement). Four studies were conducted in American populations,^20,25,29,36 9 in European populations,^{18,19,21,23,24,26,33–35} and 10 studies in Asian populations.^{8,17,22,27,28,30–32,37,38} Rodriguez et al²⁵ separately investigated 2 cohorts, and the Honululu Program had been reported earlier.¹⁷ However, different aspects of the analysis raised concerns (Table II in the online-only Data Supplement). Seven cohorts only enrolled male participants,^{8,17,21,24–27} 1 study only enrolled male smokers,²¹ and another only enrolled steelworkers.²⁷ Most studies had a follow-up duration of more than 10 years, and had a high degree of covariate adjustment, with 14 studies of “+++”,^{8,17,20,21,23,25,28,30–34,36,38} 5 of “++,”^{19,22,27,35,37} and 4 of “+.”^18,24,26,29 (Table II in the online-only Data Supplement) Cohorts and case-control studies were assessed by Newcastle–Ottawa Scale (Tables III and IV in the online-only Data Supplement). Most items had full agreement, except for 2 items. For “representativeness of the exposed cohort,” the agreement was 96% with a κ value of 0.65. For “comparability,” the agreement was 92% with a κ value of 0.75. Both items had good agreement.

Total Serum Cholesterol

High Versus Low

Seventeen studies compared the highest level with the lowest level (or referent) categories of cholesterol.^{8,18,19,21–24,26–29,31,32,34–37} Two studies compared the bottom level with a composite higher level.^17,20 Results of ICH and SAH were separately reported in 5 studies,^{8,19,21,34,37} and those of different sexes were separately assessed in 6 studies.^{18–20,22,32,34} Only 1 study investigated the difference between 2 age groups.²⁰ These separate results were initially pooled in each study using a fixed-effects model, which were further aggregated into the overall analysis, using a random-effects model. The summary RR was 0.69 (95% CI, 0.59–0.81; P<0.01; Figure 2A), with evidence of moderate heterogeneity (I²=57.6%; P<0.01). Publication bias was detected by the Egger test (P=0.03).

• Download figure
• Download PowerPoint

Dose–Response Analysis

Seventeen studies reported RRs for categorized cholesterol levels, or for per 1 mmol/L increment, including 14 prospective cohort studies,^{8,18,21–29,31,32,34} and 3 nested case-control studies.^35–37 Four studies reported both RRs for categorical levels and RRs for per 1 mmol/L increase,^27,28,34,35 and 1 study only reported RRs of per 1 mmol/L increase for 2 large cohorts.²⁵ Thus, we included 17 articles with 18 data for dose–response analysis. The number of participants in each category was not reported in 4 studies, and thus was calculated by estimation.^18,21,26,28 The summary RR was 0.85 (95% CI, 0.80–0.91; P<0.01; Figure 2B), with high heterogeneity (I²=81%; P<0.01). We detected a potentially nonlinear dose–response relationship (P<0.01; Figure 3).

• Download figure
• Download PowerPoint

Sensitivity and Subgroup Analyses

The potential sources of heterogeneity were explored by sensitivity and stratifying analyses. No significantly altered result was shown when excluding studies 1 by 1. When evaluating 1 confounding covariate, results stratified by it were separately analyzed. Studies comparing high with low levels, and studies assessing dose–response associations were explored, respectively. Overall, the inverse relationships between cholesterol levels and risk of hemorrhagic stroke were similarly significant in subgroups that were defined by sex, sample size, and follow-up duration. However, in subgroups of case-control design, SAH, “+” degree of adjustment, and end point of hemorrhagic stroke mortality, the inverse relationships were not statistically significant (Table V in the online-only Data Supplement). Besides, in the dose–-response subgroup of European population, the inverse association was not statistically significant.

High-Density Lipoprotein Cholesterol

High Versus Low

Eight studies were available, including 5 prospective cohort studies,^{21,29,31,33,34} and 3 nested case-control studies.^35,36,38 Three studies reported separate results for different types of hemorrhagic stroke, or different sexes,^21,34,35 which were aggregated together using a fixed-effects model for each study. Overall, the summary RR was 0.98 (95% CI, 0.80–1.19; P=0.81; Figure 4A), with little evidence of heterogeneity (I²=10.3%; P=0.35). Although the number of included studies was small, publication bias was not revealed by the Egger test (P=0.82).

• Download figure
• Download PowerPoint

Dose–Response Analysis

Five prospective cohort studies,^{21,29,31,33,34} and 3 nested case-control studies were included. Three studies reported separate results for different types of stroke or different sexes,^21,34,35 which were aggregated using fixed-effect model for each study. The summary RR was 1.05 (95% CI, 0.88–1.24; P=0.60; Figure 4B), with low evidence of heterogeneity (I²=35.1%; P=0.15). No potentially nonlinear dose–response relationship was detected (P=0.11; Figure 5A).

• Download figure
• Download PowerPoint

Sensitivity and Subgroup Analyses

In sensitivity analyses, no significantly altered result was shown when excluding studies 1 by 1. The stratified analyses were defined by study design, sex, stroke type, sample size, study population, degree of adjustment, and follow-up duration. Studies comparing high with low levels, and studies assessing dose–response associations were explored, respectively. In high versus low comparisons, the estimate was not significant in any subgroup. In dose–response analysis, positive relationships between HDL-C levels and risk of hemorrhagic stroke were significant in both subgroups of ICH and SAH. A marginal significance was indicated by pooling 2 studies of American population (Table VI in the online-only Data Supplement).

Low-Density Lipoprotein Cholesterol

High Versus Low

Four prospective cohort studies were included.^29–31,33 The summary RR was 0.62 (95% CI, 0.41–0.92; P=0.02; Figure 4C), with some low evidence of heterogeneity (I²=43.3%; P=0.15). In sensitivity analysis, when excluding the study by Imamura et al,³⁰ the result was more robust (RR=0.52; 95% CI, 0.37–0.72; P<0.01), with low level of heterogeneity (I²=11.1%; P=0.33). Although the number of included studies was small, publication bias was not detected by the Egger test (P=0.09).

Dose–Response Analysis

Four prospective cohort studies were identified.^29–31,33 The summary RR was 0.90 (95% CI, 0.77–1.05; P=0.18; Figure 4D), with moderate heterogeneity (I²=67%; P=0.03). In sensitivity analysis, when excluding the study by Noda et al,³¹ the summary RR was 0.93 of marginal significance (95% CI, 0.86–0.996; P=0.04) without evidence of heterogeneity (I²=0; P<0.01). We did not detect a potentially nonlinear dose–response relationship (P=0.77; Figure 5B). Too few studies precluded any meaningful subgroup analysis.

Discussion

Our findings showed a statistically significant inverse association between TC level and risk of hemorrhagic stroke. An increment of 1 mmol/L in TC concentration was associated with a 15% decreased risk of hemorrhagic stroke. Lower LDL-C concentration was also associated with a higher risk of hemorrhagic stroke. However, no significant association between HDL-C and risk of hemorrhagic stroke was indicated.

In subgroup analyses of TC, the inverse relationship seemed specifically robust for ICH but not for SAH, probably because of their different pathological mechanisms.³⁹ Statistically significant results were shown for prospective cohorts, but not for nested case-control studies. The smaller number and less strict design of case-control studies likely contributed to this discrepancy. Results for studies with adjustment degree of “+” were not significant. Seemingly, the interaction between multiple unadjusted confounders and TC leaded to weak conclusions. No significant association between TC and fatal hemorrhagic stroke was detected either. For HDL-C, the risk of ICH significantly increased for per 1 mmol/L increment. Although significant inverse relation with risk of SAH was indicated, its strength was limited by too few studies. In dose–response analysis of LDL-C, statistically significant inverse association with hemorrhagic stroke was detected by excluding 1 study.³¹

Low cholesterol may play a role in promoting arterial medial layer smooth muscle cell necrosis.³⁶ The impaired endothelium would be more susceptible to microaneurysms, which were the chief pathological finding of ICH, thereby contributing to the onset of hemorrhage.³⁹ Another mechanism was speculated especially for the association between cholesterol levels and mortality of hemorrhagic stroke.³⁶ Cholesterol levels may reflect the nutritional status of patients with ICH. Low cholesterol level may be a surrogate for nutritional deficiencies, a harbinger of low-serum albumin, or a sign of debilitating diseases,²⁰ thus being predisposed to increased stroke mortality.⁴⁰

Recent meta-analyses showed no evidence that statin therapy was associated with increased ICH.^9,10,41 However, other than reducing cholesterol levels, statins may also decrease platelet aggregation and hence thrombogenesis, which might increase the risk of bleeding and thus obscure the results.^10,42 Additionally, the follow-up durations in statin trials were usually no longer than 5 years.^43,44 Longer exposure to low cholesterol levels might be necessary to alter the integrity of cerebral vessels. Although another large meta-analysis has shown that hemorrhagic stroke mortality was likely related to lower TC levels in a few stroke subgroups,⁴ it was limited by including considerable studies published before the 1980s, and thus failed in verifying the stroke type by reliable imaging method.

In comparison, our study had strengths in conducting longer follow-up, including newer studies with reliable imaging examinations, and performing quantitative analyses. Additionally, most included studies have large sample sizes involving different general populations throughout the world. On the whole, sufficient adjustment of potential risk factors for hemorrhagic stroke was performed, and the methodological qualities were satisfying. As we included mainly prospective studies, the likelihood of selection bias and recall bias in retrospective studies was greatly reduced.

However, we were aware of several potential limitations. First, studies with significant results may be more likely to be published, and are preferentially published in English journals.¹⁵ We included only studies written in English and hence, may have missed relevant articles in non-English journals. Notably, publication bias was detected in studies of TC, which might overestimate the inverse relationship, as harmful association was more likely to be published. Second, hemorrhagic stroke is a mixed term, including ICH, SAH, subdural or epidural hemorrhages, or hemorrhagic transformation after ischemic stroke.¹⁰ Several studies only reported this composite outcome and precluded the distinction between hemorrhage subtypes.^{23,26–28,32,35,36} In fact, TC studies showed a more convincing result for ICH than SAH. Third, although adjusted estimates were reported in general, the observational design had its inherent weakness that any association may be because of the presence of inadequately measured or unmeasured residual confounding variables. For instance, the prescription of statins was not described largely,^{31,33,34,36,38} and the socioeconomic status was rarely investigated. Another major concern was that our results might be confounded by comorbidity at baseline. Because patients who died from other diseases at earlier ages could not contribute to a later risk of ICH, a survival bias possibly existed. No study excluded the first 5 or 10 years of follow-up, and still showed an association for hemorrhagic stroke. This bias may explain the protection conferred by high TC against hemorrhagic stroke, by progressively decreasing the proportion of patients with atherothrombotic disease and subsequently increasing the proportion of other cerebral vessel diseases or conditions in the elderly, which cause hemorrhagic stroke with unknown relations with cholesterols. Also, other emerging imaging confounders, including signs of leukoaraiosis, cerebral microbleeds, and multilacunes,⁴¹ were seldom mentioned. Furthermore, the number of studies was limited for HDL-C and LDL-C, and thus might be insufficient for drawing definite conclusions. Last, several studies only had the baseline data on cholesterols and no data on possible changes in serum levels during follow-up.^34,37,38

Despite these limitations, this study has notable clinical and public health implications. Our findings should never be interpreted against the well-established statin regimen, in which the potential hazards are far outweighed by the definite absolute benefits for preventing occlusive vascular disease.^43,44 Nevertheless, we highlighted that low TC and LDL-C levels seemed to be risk factors of hemorrhagic stroke, especially ICH. Our results remind clinicians to take this caution during intensive lipid-lowering therapy. Further studies are needed to investigate the underlying pathogenesis better, and identify subjects who would benefit most from lowering cholesterol without risk of hemorrhagic stroke.

Footnotes